The 20-Model Price Table
Per-request cost = 500 input tokens × input rate + 200 output tokens × output rate. Prices are USD per 1M tokens, verified 2026-08-28; DeepSeek off-peak is 50% of peak:
| Model | Rate ($/1M in, out) | Cost / request | 1M requests / mo | Coding index |
|---|---|---|---|---|
| GPT-5.6 Cyber | $12.50 / $75 | $0.02125 | $21,250 | — |
| Claude Fable 5 | $10 / $50 | $0.01500 | $15,000 | 76.5 |
| GPT-5.6 Sol | $5 / $30 | $0.00850 | $8,500 | 77.4 |
| GPT-5.5 | $5 / $30 | $0.00850 | $8,500 | 74.9 |
| Claude Opus 5 | $5 / $25 | $0.00750 | $7,500 | 78.0 |
| GPT-5.6 Terra | $2 / $12 | $0.00340 | $3,400 | 76.7 |
| Gemini 3.1 Pro | $2 / $12 | $0.00340 | $3,400 | 68.8 |
| Claude Sonnet 5 | $2 / $10 | $0.00300 | $3,000 | 71.5 |
| Grok 4.6 | $2 / $6 | $0.00220 | $2,200 | 76.8 |
| Qwen 3.8 Max | $2 / $6 | $0.00220 | $2,200 | 71.8 |
| Gemini 3.7 Flash | $1.50 / $6 | $0.00195 | $1,950 | 76.1 |
| GLM 5.3 | $1.40 / $4.40 | $0.00158 | $1,580 | 74.8 |
| DeepSeek V4 Pro | $1.32 / $3.96 | $0.00145 | $1,450 | 68.8 |
| Claude Haiku 4.5 | $1 / $5 | $0.00150 | $1,500 | 43.9 |
| Gemini 3.5 Flash-Lite | $0.30 / $2.50 | $0.00065 | $650 | 49.3 |
| Mistral Large 3 | $0.50 / $1.50 | $0.00055 | $550 | — |
| DeepSeek V4 Flash | $0.44 / $1.32 | $0.00048 | $480 | 56.2 |
| Llama 4 Maverick | $0.45 / $1.25 | $0.00048 | $480 | 16.3 |
| GPT-5.6 Luna | $0.20 / $1.20 | $0.00034 | $340 | 71.4 |
| DeepSeek V4 Flash (off-peak) | $0.22 / $0.66 | $0.00024 | $242 | 56.2 |
Three patterns jump out. Output price decides everything — GPT-5.6 Luna has the same input rate story as much pricier models but a $1.20 output rate, and it lands near the bottom. The mid-tier cluster is brutally competitive — Terra, Sonnet 5, 3.1 Pro, Grok 4.6 and Qwen 3.8 Max all sit within $0.0012 of each other. And the coding index doesn't track price: GPT-5.6 Luna scores 71.4 coding at $0.00034/request while GPT-5.6 Cyber costs 60× more with no published composite edge.
Why the Spread Is 88× (Not 5×)
The same reasoning capability costs wildly different amounts because pricing has three independent levers, and the market is still calibrating them:
Input rates collapsed to near-zero
Cached input on DeepSeek is $0.014/1M — 4% of GPT-5.6 Sol's full input. Providers compete on input because caching makes it near-free to serve; the visible input price is increasingly a marketing number.
Output rates still carry the real margin
Output is 3-25x input everywhere. The models that win on price (Luna, Flash, Flash-Lite) all pair cheap input with cheap output — that pairing is what matters, not the headline input rate.
Tier segmentation is now price-for-convenience
GPT-5.6 Sol at $5/$30 vs Terra at $2/$12: both 1.05M context, both multimodal. The flagship premium is buying reasoning depth and priority serving, not capability breadth.
What the Same Budget Buys on Each Model
Flip the table: a fixed $1,000 monthly budget at this workload size buys wildly different volumes:
That is the entire model-routing argument in one number. The routing playbook shows which tasks can take the cheap tier without quality loss — and the quality-per-dollar leaderboard shows which cheap models are actually good.
How to Price Your Own Prompt
Count your real tokens, not your word count — system prompts and tool schemas often double the input.
Decide the output budget first: output tokens are where 60-90% of your cost lives on every model in this table.
Price the worst case (flagship, uncached) and the best case (budget, off-peak, cached) — your real cost is inside that range.
Re-check quarterly: this table moves. DeepSeek, OpenAI, and Google all changed effective pricing within the last two weeks alone.
Frequently Asked Questions
How much does one prompt cost across AI models?
A 500-token prompt with a 200-token response costs between $0.00024 (DeepSeek V4 Flash off-peak) and $0.0213 (GPT-5.6 Cyber) at verified 2026-08-28 rates — an 88x spread. At 1M requests a month, that's $242 vs $21,250 for the identical workload.
Why is there such a big price difference between models?
Output tokens drive the spread: they cost 3-25x the input rate, and flagship output prices ($25-$75 per 1M) dwarf budget tiers ($0.20-$2.50). Two models with identical input rates can differ 5x per request purely on output price.
Which model is cheapest per request?
DeepSeek V4 Flash off-peak at $0.00024 per request for our 500/200 workload — and its ~97% cache-hit discount makes repeated contexts cheaper still. GPT-5.6 Luna ($0.00034) and Gemini 3.5 Flash-Lite ($0.00065) are the next cheapest.
Does cheap per-request price mean bad quality?
Not necessarily. Third-party composites (Artificial Analysis via OpenRouter) put DeepSeek V4 Pro at 53/69/50 (intelligence/coding/agentic) vs Claude Opus 5 at 63/78/59. For classification, extraction, and FAQ workloads the budget tiers are visually indistinguishable — quality-per-dollar is where they win.
How do I price my own prompt across models?
Use the prompt cost calculator with your real input/output token counts — the 20-model table below is for one representative prompt, and your mix changes the ranking (output-heavy prompts favor cheap-output models even harder).